app.py

from fastapi import FastAPI, UploadFile, File
from fastapi.middleware.cors import CORSMiddleware
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
from ultralytics import YOLO
from PIL import Image
import torch
import io

app = FastAPI()

# ✅ Allow React frontend
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# ------------------------------------------------------------
# ✅ LAZY LOAD MODELS (Important for HuggingFace Spaces)
# ------------------------------------------------------------

wound_pipe = None
yolo_model = None
tokenizer = None
chat_model = None


def load_wound_model():
    global wound_pipe
    if wound_pipe is None:
        wound_pipe = pipeline("image-classification", model="Hemg/Wound-Image-classification")
    return wound_pipe


def load_yolo_model():
    global yolo_model
    if yolo_model is None:
        yolo_model = YOLO("best.pt")   # best.pt MUST be in root
    return yolo_model


def load_llm():
    global tokenizer, chat_model
    if tokenizer is None or chat_model is None:
        LLM_ID = "google/medgemma-2b-it"     # ✅ Smaller model, fits HF
        tokenizer = AutoTokenizer.from_pretrained(LLM_ID)
        chat_model = AutoModelForCausalLM.from_pretrained(
            LLM_ID,
            device_map="cpu",               # ✅ Forces CPU mode
            torch_dtype=torch.float32
        )
    return tokenizer, chat_model


# ------------------------------------------------------------
# ✅ API 1 — Analyze Image
# ------------------------------------------------------------
@app.post("/analyze-image/")
async def analyze_image(file: UploadFile = File(...)):

    wound_pipe = load_wound_model()
    yolo_model = load_yolo_model()

    img_bytes = await file.read()
    img = Image.open(io.BytesIO(img_bytes)).convert("RGB")

    # ✅ Wound model
    wound_pred = wound_pipe(img)[0]
    wound_label = wound_pred["label"]
    wound_conf = float(wound_pred["score"])

    # ✅ YOLO detection
    yolo_out = yolo_model(img)
    if len(yolo_out[0].boxes) > 0:
        cls_id = int(yolo_out[0].boxes.cls[0])
        skin_label = yolo_model.names[cls_id]
    else:
        skin_label = "None"

    return {
        "wound_label": wound_label,
        "wound_conf": wound_conf,
        "skin_label": skin_label
    }


# ------------------------------------------------------------
# ✅ API 2 — Ask Medical Question
# ------------------------------------------------------------
@app.post("/ask-ai/")
async def ask_ai(data: dict):

    tokenizer, chat_model = load_llm()

    prompt = f"""
Detected wound: {data['wound']}
Detected skin disease: {data['skin']}
User question: {data['question']}

Respond like a medical expert.
"""

    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = chat_model.generate(**inputs, max_new_tokens=200)
    reply = tokenizer.decode(outputs[0], skip_special_tokens=True)

    return {"reply": reply}


# ✅ Start Uvicorn server
import uvicorn
if __name__ == "__main__":
    uvicorn.run("app:app", host="0.0.0.0", port=7860)